Skip to content

README.md

Latest commit

 

History

History
98 lines (80 loc) · 3.33 KB

README.md

File metadata and controls

98 lines (80 loc) · 3.33 KB

warning this is curently an ai-generated experiment so as a result should be considered slop until otherwise proven

AtomSpace GPU

An AtomSpace on the GPU.

Stores atoms as flattened Structure-of-Arrays (SoA) in VRAM so thousands of GPU threads can operate on them simultaneously. Kernels compute MI, cosine similarity, clustering, connected components, and the full learning loop (count → MI → cluster → grammar) directly on the GPU.

GpuStorageNode is the bridge that sends atoms from the CPU to the AtomSpace on the GPU and fetches results back. Same API as RocksStorageNode (disk) and CogStorageNode (network).

Supports CUDA (preferred for NVIDIA GPUs) and OpenCL (fallback for AMD/Intel). At least one must be available at build time.

Architecture

CPU (control plane)          GPU (the AtomSpace)
  Scheme / C++ code            SoA pools + kernels
        │                           │
        ├── storeAtom() ──────────→ │  receives atom
        ├── getAtom()   ←────────── │  returns results
        └── runQuery()  ──────────→ │  runs computation

The repo has two layers with different dependencies:

  • opencog/gpu/ -- The GPU AtomSpace itself. Fully standalone: pure CUDA and OpenCL with zero OpenCog dependencies. Can be built and used independently for any GPU computing project.
  • opencog/persist/gpu/ -- The GpuStorageNode bridge (~350 lines). Implements the StorageNode API so the CPU AtomSpace can send atoms to the GPU. Requires CogUtil, AtomSpace, and AtomSpace-Storage.

Building

Requires at least one of:

  • CUDA Toolkit (11.0+) for NVIDIA GPUs
  • OpenCL (1.2+ runtime and headers) for AMD/Intel/NVIDIA GPUs
mkdir build && cd build
cmake ..
make -j$(nproc)
make test
sudo make install

Usage

#include <opencog/persist/gpu/GpuStorageNode.h>

Handle hsn = as->add_node(GPU_STORAGE_NODE, std::string("gpu://:"));
GpuStorageNodePtr store = GpuStorageNodeCast(hsn);
store->open();

Handle a(createNode(CONCEPT_NODE, "someWord"));
a->setValue(truth_key(), createFloatValue({42.0, 3.7}));
store->storeAtom(a, true);
store->barrier();

Handle b(createNode(CONCEPT_NODE, "someWord"));
store->getAtom(b);  // b now has the same FloatValue

store->close();

Or from Scheme:

(use-modules (opencog) (opencog persist) (opencog persist-gpu))
(define gpu (GpuStorageNode "gpu://:"))
(cog-open gpu)
(store-atom (Concept "someWord"))
(fetch-atom (Concept "someWord"))
(cog-close gpu)

Phases

  • Phase 1 (complete): Store/fetch round-trip. 13/13 tests pass.
  • Phase 2 (planned): runQuery() -- execute Atomese expressions on GPU.
  • Phase 3 (planned): JIT compilation -- fused kernels from AST.

Known Limitations

  • GPU storage is volatile -- data is lost when the connection closes.
  • Only truth_key() values are stored (GPU has fixed SoA layout).
  • CUDA backend requires Compute Capability 5.0+ (Maxwell or newer).